1. Field of the Invention
The present invention relates to improvements in data processing systems. More particularly, the invention is directed to eliminating performance bottlenecks and reducing system size and cost by increasing the memory, processing, and I/O capabilities that can be integrated into a monolithic region.
2. Description of Prior Art
Early computer circuits were made of separate components wired together on a macroscopic scale. The integrated circuit combined all circuit components (resistors, capacitors, transistors, and conductors) onto a single substrate, greatly decreasing circuit size and power consumption, and allowing circuits to be mass produced already wired together. This mass production of completed circuitry initiated the astounding improvements in computer performance, price, power and portability of the past few decades. But lithographic errors have set limits on the complexity of circuitry that can be fabricated in one piece without fatal flaws. To eliminate these flaws large wafers of processed substrate are diced into chips so that regions with defects can be discarded. Improvements in lithography allow continually increasing levels of integration on single chips, but demands for more powerful and more portable systems are increasing faster still.
Portable computers using single-chip processors can be built on single circuit boards today, but because lithographic errors limit the size and complexity of today's chips, each system still requires many separate chips. Separate wafers of processor, memory, and auxiliary chips are diced into their component chips, a number of which are then encapsulated in bulky ceramic packages and affixed to an even bulkier printed circuit board to be connected to each other, creating a system many orders of magnitude bigger than its component chips. Using separate chips also creates off-chip data flow bottlenecks because the chips are connected on a macroscopic rather than a microscopic scale, which severely limits the number of interconnections. Macroscopic inter-chip connections also increase power consumption. Furthermore, even single board systems use separate devices external to that board for system input and output, further increasing system size and power consumption. The most compact systems thus suffer from severe limits in battery life, display resolution, memory, and processing power.
Reducing data traffic across the off-chip bottleneck and increasing processor-to-memory connectivity through adding memory to processor chips is known in the art. Both Intel's new Pentium (tm) processor and IBM/Motorola/Apple's PowerPC (tm) 601 processor use 256-bit-wide data paths to small on-chip cache memories to supplement their 64-bit wide paths to their systems' external-chip main memories ("RISC Drives PowerPC", BYTE, August 1993, "Intel Launches a Rocket in a Socket", BYTE, May 1993). Chip size limits, however, prevent the amount of on-chip memory from exceeding a tiny fraction of the memory used in a whole system.
Parallel computer systems are well known in the art. IBM's 3090 mainframe computers, for example, use parallel processors sharing a common memory. While such shared memory parallel systems do remove the von Neumann uniprocessor bottleneck, the funneling of memory access from all the processors through a single data path rapidly reduces the effectiveness of adding more processors. Parallel systems that overcome this bottleneck through the addition of local memory are also known in the art. U.S. Pat. No. 5,056,000, for example, discloses a system using both local and shared memory, and U.S. Pat. No. 4,591,981 discloses a local memory system where each "local memory processor" is made up of a number of smaller processors sharing that "local" memory. But in these systems the local processor/memory clusters contain many separate chips, and while each processor has its own local input and output, that input and output is done through external devices. This requires complex macroscopic (and hence off-chip-bottleneck-limited) connections between the processors and external chips and devices, which rapidly increases the cost and complexity of the system as the number of processors is increased.
Massively parallel computer systems are also known in the art. U.S. Pat. Nos. 4,622,632, 4,720,780, 4,873,626, and 4,942,517, for instance, disclose examples of systems comprising arrays of processors where each processor has its own memory. While these systems do remove the von Neumann uniprocessor bottleneck and the multi-processor memory bottleneck for parallel applications, the processor/memory connections and the interprocessor connections are still limited by the off-chip data path bottleneck. Also, the output of the processors is still gathered together and funneled through a single data path to reach a given external output device, which creates an output bottleneck that limits the usefulness of such systems for output-intensive tasks. The use of external input and output devices further increases the size, cost and complexity of the overall systems.
Even massively parallel computer systems where separate sets of processors have separate paths to I/O devices, such as those disclosed in U.S. Pat. Nos. 4,591,980, 4,933,836 and 4,942,517 and Thinking Machines Corp.'s CM-5 Connection Machine (tm), rely on connections to external devices for their input and output ("Machines from the Lunatic Fringe", TIME, Nov. 11, 1991). Having each processor set connected to an external I/O device also necessitates having a multitude of connections between the processor array and the external devices, thus greatly increasing the overall size, cost and complexity of the system. Furthermore, output from multiple processors to a single output device, such as an optical display, is still gathered together and funneled through a single data path to reach that device. This creates an output bottleneck that limits the usefulness of such systems for display-intensive tasks.
Multi-processor chips are also known in the art. U.S. Pat. No. 5,239,654, for example, calls for "several" parallel processors on an image processing chip. Even larger numbers of processors are possible--Thinking Machines Corp.'s original CM-1 Connection Machine, for example, used 32 processors per chip to reduce the numbers of separate chips and off-chip connections needed for (and hence the size and cost of) the system as a whole (U.S. Pat. No. 4,709,327). The chip-size limit, however, forces a severe trade-off between number and size of processors in such architectures; the CM-1 chip used 1-bit processors instead of the 8-bit to 32-bit processors in common use at that time. But even for massively parallel tasks, trading one 32-bit processor per chip for 32 one-bit processors per chip does not produce any performance gains except for those tasks where only a few bits at a time can be processed by a given processor. Furthermore, these non-standard processors do not run standard software, requiring everything from operating systems to compilers to utilities to be re-written, greatly increasing the expense of programming such systems. Newer massively parallel systems such as the CM-5 Connection Machine use standard 32-bit full-chip processors instead of multi-processor chips.
Input arrays are also known in the art. State-of-the-art video cameras, for example, use arrays of charge-coupled devices (CCD's) to gather parallel optical inputs into a single data stream. Combining an input array with a digital array processor is disclosed in U.S. Pat. No. 4,908,751, with the input array and processor array being separate devices and the communication between the arrays being shown as row-oriented connections, which would relieve but not eliminate the input bottleneck. Input from an image sensor to each processing cell is mentioned as an alternative input means in U.S. Pat. No. 4,709,327, although no means to implement this are taught. Direct input arrays that do analog filtering of incoming data have been pioneered by Carver Mead, et al., ("The Silicon Retina", Scientific American, May 1991). While this direct-input/analog-filtering array does eliminate the input bottleneck to the array, these array elements are not suitable for general data processing. All these arrays also lack direct output means and hence do not overcome the output bottleneck, which is far more critical in most real-world applications. The sizes of these arrays are also limited by lithographic errors, so systems based on such arrays are subjected to the off-chip data flow bottleneck. Reliance on connections to external output devices also increases the overall size, cost and complexity of those systems.
Output arrays where each output element has its own transistor are also known in the art and have been commercialized for flat-panel displays, and some color displays use display elements with one transistor for each color. Since the output elements cannot add or subtract or edit-and-pass-on a data stream, such display elements can do no data decompression or other processing, so the output array requires a single uncompressed data stream, creating a band-width bottleneck as array size increases. These output arrays also have no defect tolerance, so every pixel must be functional or an obvious "hole" will show up in the array. This necessity for perfection creates low yields and high costs for such displays.
Systems that use wireless links to communicate with external devices are also known in the art. Cordless data transmission devices, including keyboards and mice, hand-held computer to desk-top computer data links, remote controls, and portable phones are increasing in use every day. But increased use of such links and increases in their range and data transfer rates are all increasing their demands for bandwidth. Some electromagnetic frequency ranges are already crowded, making this transmission bottleneck increasingly a limiting factor. Power requirements also limit the range of such systems and often require the transmitter to be physically pointed at the receiver for reliable transmission to occur.
Integrated circuits fabricated from amorphous and polycrystalline silicon, as opposed to crystalline silicon, are also known in the art. These substrates, though, are far less consistent and have lower electron mobility, making it difficult to fabricate fast circuits without faults. Since circuit speed and lithographic errors cause significant bottlenecks in today's computers, the slower amorphous and polycrystalline silicon integrated circuits have not been competitive with crystalline silicon in spite of their potentially lower fabrication costs.
Fault-tolerant architectures are also known in the art. The most successful of these are the spare-line schemes used in memory chips. U.S. Pat. Nos. 3,860,831 and 4,791,319, for example, disclose spare-line schemes suitable for such chips. In practice, a 4 megabit chip, for example, might nominally have 64 cells each with 64k active bits of memory in a 256.times.256 bit array, while each cell physically has 260 bits by 260 bits connected in a manner that allows a few errors per cell to be corrected by substituting spare lines, thus saving the cell. This allows a finer lithography to be used, increasing the chip's memory density and speed. Since all bits in a memory chip have the same function, such redundancy is relatively easy to implement for memory. Processors, however, have large numbers of circuits with unique functions (often referred to in the art as random logic circuits), and a spare circuit capable of replacing one kind of defective circuit cannot usually replace a different kind, making these general spare-circuit schemes impractical for processors.
Redundancy schemes that handle random logic circuits by replicating every circuit are also known in the art. These incorporate means for selecting the output of a correctly functioning copy of each circuit and ignoring or eliminating the output of a faulty copy. Of these replication schemes, circuit duplication schemes, as exemplified by U.S. Pat. Nos. 4,798,976 and 5,111,060, use the least resources for redundancy, but provide the least protection against defects because two defective copies of a given circuit (or a defect in their joint output line) still creates an uncorrectable defect. Furthermore, it is necessary to determine which circuits are defective so that they can be deactivated. Many schemes therefore add a third copy of every circuit so that a voting scheme can automatically eliminate the output of a single defective copy. This, however, leads to a dilemma: When the voting is done on the output of large blocks of circuitry, there is a significant chance that two out of the three copies will have defects, but when the voting is done on the output of small blocks of circuitry, many voting circuits are needed, increasing the likelihood of errors in the voting circuits themselves! Ways to handle having two defective circuits out of three (which happens more frequently than the 2 defects out of 2 problem that the duplication schemes face) are also known. One tactic is to provide some way to eliminate defective circuits from the voting, as exemplified by U.S. Pat. No. 4,621,201. While this adds a diagnostic step to the otherwise dynamic voting process, it does allow a triplet with two defective members to still be functional. Another tactic, as exemplified by U.S. Pat. Nos. 3,543,048 and 4,849,657, calls for N-fold replication, where N can be raised to whatever level is needed to provide sufficient redundancy. Not only is a large N an inefficient use of space, but it increases the complexity of the voting circuits themselves, and therefore the likelihood of failures in them. This problem can be reduced somewhat, although not eliminated, by minimizing the complexity of the voting circuits, as U.S. Pat. No. 4,617,475 does through the use of an analog differential transistor added to each circuit replicate, allowing a single analog differential transistor to do the voting regardless of how many replicates of the circuit there are. Yet another tactic is to eliminate the "voting" by replicating circuits at the gate level to build the redundancy into the logic circuit themselves. U.S. Pat. No. 2,942,193, for example, calls for quadruplication of every circuit, and uses an interconnection scheme that eliminates faulty signals within two levels of where they originate. While this scheme can be applied to integrated circuits (although it predates them considerably), it requires four times as many gates, each with twice as many inputs, as equivalent non-redundant logic, increasing the circuit area and power requirements too much to be practical. All these N-fold redundancy schemes also suffer from problems where if the replicates are physically far apart, gathering the signals requires extra wiring, creating propagation delays, while if the replicates are close together, a single large lithographic error can annihilate the replicates en masse, thus creating an unrecoverable fault.
Cell-based fault-tolerant architectures are also known in the art. U.S. Pat. Nos. 3,913,072 and 5,203,005, for example, both disclose fault-tolerant schemes that connect whole wafers of cells into single fault-free cell chains, even when a significant number of the individual cells are defective. The resulting one-dimensional chains, however, lack the direct addressability needed for fast memory arrays, the positional regularity of array cells needed for I/O arrays, and the two-dimensional or higher neighbor-to-neighbor communication needed to efficiently handle most parallel processing tasks. This limits the usefulness of these arrangements low or medium performance memory systems and to tasks dominated by one-dimensional or lower connectivity, such as sorting data. U.S. Pat. No. 4,800,302 discloses a global address bus based spare cell scheme that doesn't support direct cell-to-cell connections at all, requiring all communications between cells to be on the global bus. Addressing cells through a global bus has significant drawbacks; it does not allow parallel access of multiple cells, and comparing the cell's address with an address on the bus introduces a delay in accessing the cell. Furthermore, with large numbers of cells it is an inefficient user of power; in order for N cells to determine whether they are being addressed, each must check a minimum of log2(N) address bits (in binary systems), so an address signal requires enough power to drive N*log2(N) inputs. This is a high price in a system where all intercell signals are global.
Even cell-based fault-tolerant architectures that support two-dimensional connectivity are known in the art. U.S. Pat. No. 5,065,308 discloses a cell array that can be organized into a series of fault-free linear cell chains or a two-dimensional array of fault-free cells with neighbor-to-neighbor connections. Several considerations, however, diminish its applicability to large high-performance array at all but the lowest defect densities. While the cells can be addressed through their row and column connections IPN.fwdarw.OPS and IPE.fwdarw.OPW, this addressing is not direct in that a signal passing from West to East encounters two 3-input gates per cell, (even assuming zero-delay passage through the processor itself). Thus while large cells create high defect rates, small cells sizes create significant delays in the propagation of signals across the array. Consider, for example, a wafer with 1 defect per square centimeter, which is reasonable for a leading edge production technology. On a 5" wafer an 80 square centimeter rectangular array can be fabricated. Now consider what size cells might be suitable. With an 8 by 10 array of 1 cm square cells (less than half the size of a Pentium chip) the raw cell yield would be around 30%, or an average of 24 or 25 good cells. Only when every single column had at least one good cell, and that spaced by at most one row from the nearest good cell in each of the neighboring columns, could even a single 1.times.8 fault-free cell "array" could be formed. This should happen roughly 10% of the time, for an abysmal overall 1% array cell yield. With wafer scale integration, however, smaller cell sizes are useful as the cells do not have to be diced and reconnected. As cell size decreases, yields grow rapidly, but the propagation delays grow, too. With 5 mm square cells a 16.times.20 raw cell array would fit, and the raw cell yield would be almost 75%, so most arrays would have around 240 good cells. While an average column would have 15 good cells, it is the column with the fewest good cells that determine the number of rows in the final array. This would typically be 10 or 11 rows, creating 16.times.10 or 16.times.11 arrays. This would be a 50%-55% array cell yield, which is quite reasonable. But row-addressing signals propagated across the array would pass sequentially through up to 30 gates, creating far too long a delay for high-performance memory systems.
This interconnection scheme also has problems when used for processing cells, although it is targeted for that use. The cell bypassing scheme does support two-dimensional neighbor-to-neighbor connectivity, and could support a column-oriented bus for each column, but it cannot support a corresponding row-oriented bus without the 2-gate-per-cell delay. Three dimensional connectivity could be accomplished only by extending the bypass scheme to physically three dimensional arrays, which cannot be made with current lithography, and higher-dimensional connectivities such as hyper-cube connectivity are out of the question. Even for two-dimensional neighbor-to-neighbor connectivity, this scheme has certain drawbacks. While the row-oriented neighbor-to-neighbor connections never span a distance larger than one diagonal cell-center to cell-center, column-oriented neighbor-to-neighbor connections can be forced to span several defective or inactive cells. All intercell timing and power considerations must take into account the maximum capacitances and resistances likely to be encountered on such a path. This scheme also shifts the position of every cell in the entire rest of the column (relative to its same-logical-row neighbors) for each defective cell that is bypassed, which propagates the effects of each defective cell far beyond the neighborhood of the defect. This multi-cell shift also prevents this scheme from being useful in arrays where physical position of array cells is important, such as direct input or output cell arrays.